@xopcai/xopc 0.0.244 → 0.0.245

package/dist/src/agent/orchestration/loop-guard.js

@@ -1,105 +1,20 @@
-import { createLogger } from "../../utils/logger/index.js";
-import { init_logger } from "../../utils/logger.js";
+import { createHash } from "node:crypto";
 //#region src/agent/orchestration/loop-guard.ts
-init_logger();
-const log = createLogger("LoopGuard");
-const DEFAULT_CONFIG = {
-	softThreshold: 2,
-	hideThreshold: 3
-};
-/**
-* Stable fingerprint for tool call comparison.
-* Recursively sorts object keys; truncates long string values.
-*/
-function stableJson(value) {
-	return JSON.stringify(value, (_key, item) => {
-		if (typeof item === "object" && item !== null && !Array.isArray(item)) return Object.fromEntries(Object.entries(item).sort(([a], [b]) => a.localeCompare(b)));
-		if (typeof item === "string" && item.length > 200) return `${item.slice(0, 80)}…[${item.length}]`;
-		return item;
-	});
-}
 function fingerprint(call) {
-	const result = call.resultPreview === void 0 ? "" : `::result=${stableJson(call.resultPreview)}`;
-	return `${call.name}::${stableJson(call.params)}${result}`;
-}
-/**
-* Analyze recent tool calls and produce loop guard result.
-*
-* Pure function — no side effects, no shared state.
-*/
-function detectToolLoops(recentCalls, config = {}) {
-	const { softThreshold, hideThreshold } = {
-		...DEFAULT_CONFIG,
-		...config
-	};
-	if (recentCalls.length === 0) return {
-		injection: null,
-		hiddenTools: /* @__PURE__ */ new Set()
-	};
-	const consecutiveGroups = findConsecutiveRepeats(recentCalls);
-	const frequencyMap = countFrequency(recentCalls);
-	const warnings = [];
-	const hiddenTools = /* @__PURE__ */ new Set();
-	for (const group of consecutiveGroups) if (group.count >= hideThreshold) {
-		hiddenTools.add(group.toolName);
-		warnings.push(`Tool '${group.toolName}' has been called ${group.count} times consecutively with identical arguments and is now unavailable. Use a completely different approach.`);
-		log.warn({
-			tool: group.toolName,
-			consecutiveCount: group.count
-		}, "Loop guard: hiding tool from LLM");
-	} else if (group.count >= softThreshold) {
-		warnings.push(`You have called '${group.toolName}' ${group.count} times with the same arguments. This is not producing new information. Do NOT call it again with the same arguments — try a different approach.`);
-		log.debug({
-			tool: group.toolName,
-			consecutiveCount: group.count
-		}, "Loop guard: soft warning injected");
-	}
-	for (const [fp, count] of frequencyMap) {
-		const toolName = fp.split("::")[0];
-		if (count >= hideThreshold + 1 && !hiddenTools.has(toolName)) {
-			if (!consecutiveGroups.some((g) => g.toolName === toolName && g.count >= softThreshold)) warnings.push(`Tool '${toolName}' has been called ${count} times this turn with the same arguments (interleaved). Stop repeating this pattern.`);
-		}
-	}
-	return {
-		injection: warnings.length > 0 ? `⚠️ LOOP DETECTION:\n${warnings.join("\n")}\n\nYou MUST take a different approach. Do not repeat the same tool calls.` : null,
-		hiddenTools
-	};
-}
-/**
-* Find consecutive runs of identical (tool+params) calls from the end of history.
-* Returns all groups with count >= 2 (most recent first).
-*/
-function findConsecutiveRepeats(calls) {
-	const groups = [];
-	let index = calls.length - 1;
-	while (index >= 0) {
-		const current = calls[index];
-		const fp = fingerprint(current);
-		let count = 1;
-		while (index - count >= 0) {
-			const prev = calls[index - count];
-			if (fingerprint(prev) !== fp) break;
-			count++;
-		}
-		if (count >= 2) groups.push({
-			toolName: current.name,
-			fingerprint: fp,
-			count
-		});
-		index -= count;
-	}
-	return groups;
+	const value = JSON.stringify(call, (_key, item) => item && typeof item === "object" && !Array.isArray(item) ? Object.fromEntries(Object.entries(item).sort(([a], [b]) => a.localeCompare(b))) : item);
+	return createHash("sha256").update(value).digest("hex");
 }
-/**
-* Count how many times each (tool+params) fingerprint appears.
-*/
-function countFrequency(calls) {
-	const map = /* @__PURE__ */ new Map();
-	for (const call of calls) {
-		const fp = fingerprint(call);
-		map.set(fp, (map.get(fp) ?? 0) + 1);
-	}
-	return map;
+/** Warn on a stalled tail, without removing tools needed to recover or verify a repair. */
+function detectToolLoops(recentCalls) {
+	const calls = recentCalls.slice(-12);
+	const last = calls.at(-1);
+	if (!last) return { injection: null };
+	const hashes = calls.map(fingerprint);
+	let count = 0;
+	for (let index = hashes.length - 1; index >= 0 && hashes[index] === hashes.at(-1); index--) count++;
+	const alternating = hashes.length >= 6 && hashes.slice(-6, -3).every((hash, index) => hash === hashes.at(-3 + index));
+	if (count < 2 && !alternating) return { injection: null };
+	return { injection: `Repeated tool calls are producing no new evidence (${last.name}, ${count >= 2 ? `${count} identical calls` : "repeated cycle"}). Change the query or fix the cause before retrying. For a running job use managed_job wait. Tools remain available for a different argument or a changed workspace.` };
 }
 //#endregion
 export { detectToolLoops };

package/dist/src/agent/prompt/sections/tooling.js

@@ -34,8 +34,7 @@ const CORE_TOOL_ORDER = [
 	"text_to_speech",
 	"image",
 	"image_generate",
-	"create_share",
-	"execute_code"
+	"create_share"
 ];
 const CORE_TOOL_SUMMARIES = {
 	read_file: "Read targeted file contents before editing",
@@ -72,8 +71,7 @@ const CORE_TOOL_SUMMARIES = {
 	text_to_speech: "Send a standalone voice message",
 	image: "Analyze an image with the configured vision model",
 	image_generate: "Generate images with the configured image model",
-	create_share: "Create a shareable link for workspace files",
-	execute_code: "Run code in a sandbox with a restricted tool subset"
+	create_share: "Create a shareable link for workspace files"
 };
 function buildToolingSection(params) {
 	const rawToolNames = (params.toolNames ?? []).map((tool) => tool.trim()).filter(Boolean);

package/dist/src/agent/sandbox/types.d.ts

@@ -1,8 +1,8 @@
 /**
- * Sandbox type definitions for Phase 1 process-level isolation.
+ * Preflight command/path policy. OS isolation is provided by the command runtime.
  */
 /** Controls how strictly the sandbox enforces isolation. */
-export type SandboxMode = 'off' | 'basic' | 'container';
+export type SandboxMode = 'off' | 'basic';
 /** Result of path validation against the sandbox policy. */
 export type PathValidationResult = {
     allowed: boolean;

package/dist/src/agent/sandbox/types.js

@@ -3,7 +3,7 @@
 const DEFAULT_SANDBOX_CONFIG = {
 	mode: "basic",
 	allowedRoots: [],
-	maxExecutionTimeMs: 3e5,
+	maxExecutionTimeMs: 240 * 6e4,
 	maxOutputBytes: 50 * 1024
 };
 //#endregion

package/dist/src/agent/service/direct-turn-helpers.js

@@ -1,6 +1,6 @@
-import { buildTaskExecutionDirective } from "../../tasks/task-context-assembler.js";
 import { resolveImageHandlingStrategy } from "../image/vision-detection.js";
 import { hydrateUserTurnForLlm } from "../inbound/attachment-pipeline.js";
+import { buildTaskExecutionDirective } from "../../tasks/task-context-assembler.js";
 import { prependAgentContext } from "../../user-context/planner.js";
 import { extractAgentUserPlainText } from "../memory/user-message-text.js";
 import { shouldSkipResetOverlapCommand } from "../../session/reset-triggers.js";

package/dist/src/agent/service/process-direct-streaming.js

@@ -1,4 +1,5 @@
 import { readAgentMessageContent } from "../memory/agent-message-access.js";
+import { abortEmbeddedRun } from "../embedded/runs.js";
 import { clearPendingTranscriptUserMessage, setPendingTranscriptUserMessage } from "../inbound/attachment-pipeline.js";
 import { formatAgentRunErrorForClient } from "../client-error-format.js";
 import { resolveConfiguredActivityDetailDefault, resolveEffectiveReasoningLevel } from "../../session/thinking-resolve.js";
@@ -6,7 +7,6 @@ import { initSessionTurn } from "../../session/init-session-turn.js";
 import "../../session/index.js";
 import { mergeSttConfigFromAppConfig } from "../../voice/stt/config.js";
 import "../../voice/stt/index.js";
-import { abortEmbeddedRun } from "../embedded/runs.js";
 import { isVoiceLikeAttachment, mergeVoiceTranscriptsIntoUserText } from "../../channels/attachments/voice-stt-webchat.js";
 import { applyReasoningVisibility } from "../streaming/reasoning-visibility.js";
 import { isSessionSourceBinding, summarizeSourceContext } from "../source-context/types.js";

package/dist/src/agent/service.d.ts

@@ -38,8 +38,6 @@ export declare class AgentService {
     private toolUsageAnalyzer;
     private toolChainTracker;
     private errorPatternMatcher;
-    private selfVerifyMiddleware;
-    private turnDiffTracker;
     private contextMiddleware;
     private messageRouter;
     private commandHandler;

package/dist/src/agent/service.js

@@ -3,10 +3,10 @@ import { init_logger } from "../utils/logger.js";
 import { getUserProfile } from "../storage/sqlite/user-context-repository.js";
 import { parseNoteAttachmentTarget } from "../notes/attachment-ref.js";
 import { getAgentDefaultModelRef, init_schema } from "../config/schema.js";
-import { applyConfigOverrides } from "../config/runtime-overrides.js";
 import { init_agent_scope, resolveAgentHomeDir, resolveDefaultAgentId } from "./agent-scope.js";
-import { extractProfileAgentId } from "../config/agent-profile.js";
 import { init_providers, resolveModel } from "../providers/index.js";
+import { applyConfigOverrides } from "../config/runtime-overrides.js";
+import { extractProfileAgentId } from "../config/agent-profile.js";
 import { extractTextContent } from "./context/workspace.js";
 import { effectiveWorkspacePathForSession } from "../session/session-workspace.js";
 import { getProjectForSession } from "../projects/workspace.js";
@@ -15,10 +15,11 @@ import { removeSkillsLockEntry } from "./skills/hub-lock.js";
 import { installSkillFromZip } from "./skills/managed-store.js";
 import { pullSkillFromSource } from "./skills/hub-pull.js";
 import { resolveAgentTurnTimeoutMs } from "./orchestration/run-agent-turn-with-timeout.js";
-import { resolveResponseLanguageForSession } from "./prompt/response-language.js";
-import { normalizeSkillInstallTarget } from "./skills/install-target.js";
 import { onSessionTranscriptUpdate } from "../session/transcript-events.js";
 import { persistInboundAttachments } from "../channels/attachments/inbound-persist.js";
+import { runXopcEmbeddedTurn } from "./embedded/run-turn.js";
+import { resolveResponseLanguageForSession } from "./prompt/response-language.js";
+import { normalizeSkillInstallTarget } from "./skills/install-target.js";
 import { AgentManager } from "./agent-manager.js";
 import { maybeRefineSessionTitleWithLlm, maybeSetProvisionalSessionTitle } from "../session/session-title.js";
 import { resolveCompactionPolicy } from "./memory/compaction-policy.js";
@@ -38,13 +39,10 @@ import { SystemReminder } from "./prompt/system-reminder.js";
 import { ToolUsageAnalyzer } from "./tools/usage-analyzer.js";
 import { ToolChainTracker } from "./tools/chain-tracker.js";
 import { ErrorPatternMatcher } from "./tools/error-pattern-matcher.js";
-import { SelfVerifyMiddleware } from "./middleware/self-verify.js";
 import { ContextMiddleware } from "./middleware/context.js";
 import "./middleware/index.js";
-import { TurnDiffTracker } from "./coding/turn-diff-tracker.js";
 import { LifecycleManager } from "./lifecycle/manager.js";
 import "./lifecycle/index.js";
-import { runXopcEmbeddedTurn } from "./embedded/run-turn.js";
 import { runWithEmbeddedExecutionSession } from "./embedded/execution-context.js";
 import { MessageRouter } from "./messaging/message-router.js";
 import { CommandHandler } from "./messaging/command-handler.js";
@@ -106,8 +104,6 @@ var AgentService = class {
 	toolUsageAnalyzer;
 	toolChainTracker;
 	errorPatternMatcher;
-	selfVerifyMiddleware;
-	turnDiffTracker;
 	contextMiddleware;
 	messageRouter;
 	commandHandler;
@@ -242,19 +238,16 @@ var AgentService = class {
 			onSkillsUpdated: config.onSkillsUpdated,
 			installSkillFromSource: (opts) => this.installSkillFromSource(opts),
 			installSkillFromMarketplace: (opts) => this.installSkillFromMarketplace(opts),
-			isWorkspaceTrusted: config.isWorkspaceTrusted,
-			getSelfVerifyPromptContext: (sessionKey, agentId) => [this.selfVerifyMiddleware.getPendingVerificationContext(sessionKey, agentId), this.turnDiffTracker.buildFinalGuardContext(sessionKey)].filter(Boolean).join("\n\n")
+			isWorkspaceTrusted: config.isWorkspaceTrusted
 		});
 		this.agentEventHandler = new AgentEventHandler({
 			errorTracker: this.errorTracker,
 			requestLimiter: this.requestLimiter,
 			lifecycleManager: this.lifecycleManager,
 			toolChainTracker: this.toolChainTracker,
-			selfVerifyMiddleware: this.selfVerifyMiddleware,
 			systemReminder: this.systemReminder,
 			toolUsageAnalyzer: this.toolUsageAnalyzer,
-			errorPatternMatcher: this.errorPatternMatcher,
-			turnDiffTracker: this.turnDiffTracker
+			errorPatternMatcher: this.errorPatternMatcher
 		});
 		this.agentEventHandler.registerListener("message_end", (event, context) => {
 			const message = event.message;
@@ -438,13 +431,6 @@ var AgentService = class {
 			maxTotalFailures: 5,
 			resetOnTurnEnd: true
 		});
-		this.selfVerifyMiddleware = new SelfVerifyMiddleware({
-			maxEditsPerFile: 5,
-			enablePreCompletionCheck: true,
-			minTurnsForVerification: 4,
-			resetOnVerification: true
-		});
-		this.turnDiffTracker = new TurnDiffTracker();
 		this.requestLimiter = new RequestLimiter({
 			maxRequestsPerTurn: 50,
 			warnThreshold: .8,
@@ -836,10 +822,7 @@ var AgentService = class {
 	handleSessionEvent(sessionKey, event) {
 		const currentContext = this.sessionContextManager.getContext();
 		if (!currentContext) return;
-		if (currentContext.sessionKey !== sessionKey) {
-			this.agentEventHandler.handle(event, currentContext);
-			return;
-		}
+		if (currentContext.sessionKey !== sessionKey) return;
 		if (event.type === "message_update") {
 			const msgEvent = event;
 			if (msgEvent.message?.role === "assistant") {

package/dist/src/agent/session/session-inspector.js

@@ -1,8 +1,8 @@
 import { createLogger } from "../../utils/logger/index.js";
 import { init_logger } from "../../utils/logger.js";
 import { init_workspace_defaults, resolveDefaultAgentWorkspaceDir } from "../../config/workspace-defaults.js";
-import { resolveEffectiveAgentProfileForSession } from "../../config/agent-profile.js";
 import { init_providers, resolveModel } from "../../providers/index.js";
+import { resolveEffectiveAgentProfileForSession } from "../../config/agent-profile.js";
 import { getExecutionEnvironmentForSession } from "../../execution-environments/subject.js";
 import { effectiveWorkspacePathForSession, projectWorkspacePath } from "../../session/session-workspace.js";
 import { getProjectForSession } from "../../projects/workspace.js";

package/dist/src/agent/skills/hub-lock.js

@@ -1,5 +1,5 @@
-import { init_paths, resolveSkillsLockPath } from "../../config/paths.js";
 import { init_write_file_atomic, writeTextAtomicSync } from "../../infra/write-file-atomic.js";
+import { init_paths, resolveSkillsLockPath } from "../../config/paths.js";
 import { existsSync, readFileSync } from "node:fs";
 //#region src/agent/skills/hub-lock.ts
 /**

package/dist/src/agent/tasks/task-judge-service.d.ts

@@ -1,6 +1,7 @@
 import type { TaskJudgment } from '@xopcai/gateway-contract';
 import { type Config } from '../../config/schema.js';
 import type { SessionStore } from '../../session/store.js';
+import type { TranscriptStoredRow } from '../../session/session-context-for-llm.js';
 import type { ModelManager } from '../models/index.js';
 export interface TaskTurnCompletion {
     sessionKey: string;
@@ -19,6 +20,12 @@ export interface TaskJudgeDecision {
     judgment: TaskJudgment;
 }
 export declare function parseTaskJudgeDecision(raw: string, criteriaCount: number): TaskJudgeDecision;
+export declare function codingCompletionEvidence(rows: readonly TranscriptStoredRow[]): {
+    allowed: boolean;
+    workspace?: string;
+    revision?: string;
+    evidence?: unknown;
+};
 export declare class TaskJudgeService {
     #private;
     private readonly options;

package/dist/src/agent/tasks/task-judge-service.js

@@ -1,14 +1,15 @@
 import { createLogger } from "../../utils/logger/index.js";
 import { init_logger } from "../../utils/logger.js";
+import { extractAssistantText, getAssistantMessageErrorReason, stripCodeFences } from "../../providers/model-response.js";
 import { getAgentDefaultModelRef, init_schema } from "../../config/schema.js";
-import { resolveEffectiveAgentProfileForSession } from "../../config/agent-profile.js";
 import { getApiKey, init_providers, resolveModel } from "../../providers/index.js";
+import { completeWithResolvedCredentials } from "../../providers/model-call.js";
+import { resolveEffectiveAgentProfileForSession } from "../../config/agent-profile.js";
 import { TaskRepository } from "../../tasks/task-repository.js";
 import { TaskConversationRepository } from "../../tasks/task-conversation-repository.js";
 import { TaskRunRepository } from "../../tasks/task-run-repository.js";
-import { completeWithResolvedCredentials } from "../../providers/model-call.js";
+import { readWorkspaceRevision } from "../coding/workspace-revision.js";
 import { TaskApplicationService } from "../../tasks/task-application-service.js";
-import { extractAssistantText, getAssistantMessageErrorReason, stripCodeFences } from "../../providers/model-response.js";
 //#region src/agent/tasks/task-judge-service.ts
 init_schema();
 init_providers();
@@ -51,6 +52,21 @@ function parseTaskJudgeDecision(raw, criteriaCount) {
 		}
 	};
 }
+function codingCompletionEvidence(rows) {
+	const entries = rows;
+	const start = entries.findLastIndex((row) => row.type === "custom" && row.customType === "coding_run_started");
+	const final = entries.findLastIndex((row) => row.type === "custom" && row.customType === "coding_verification");
+	const checkpoint = entries[Math.max(start, final)]?.data;
+	if (!checkpoint?.required) return { allowed: true };
+	if (start > final) return { allowed: false };
+	const evidence = Array.isArray(checkpoint.evidence) ? checkpoint.evidence : [];
+	return {
+		allowed: checkpoint.changed !== true || evidence.some((item) => item.kind === "check" && item.status === "passed") && evidence.some((item) => item.kind === "diff-review" && item.status === "passed") && evidence.every((item) => item.status === "passed"),
+		evidence,
+		workspace: typeof checkpoint.workspace === "string" ? checkpoint.workspace : void 0,
+		revision: typeof checkpoint.revision === "string" ? checkpoint.revision : void 0
+	};
+}
 function resolveJudgeModel(config, sessionKey, runtimeModel) {
 	if (config) {
 		const profile = resolveEffectiveAgentProfileForSession(config, sessionKey);
@@ -94,12 +110,14 @@ var TaskJudgeService = class {
 			return;
 		}
 		const history = compactHistory(await this.options.sessionStore.loadMessages(payload.sessionKey));
+		const proof = codingCompletionEvidence(await this.options.sessionStore.loadTranscriptRows(payload.sessionKey));
 		const criteria = task.contract.acceptanceCriteria;
 		const request = {
 			role: "user",
 			content: [
 				"You are an independent task verifier. Be strict and evidence-driven.",
-				"Mark a criterion complete only when the transcript or latest response directly proves it.",
+				"Mark a criterion complete only when observed tool results or user-provided facts prove it. An assistant completion claim is not verification evidence.",
+				`Runtime coding evidence: ${JSON.stringify(proof)}`,
 				"needsUser is true only when a specific decision, permission, credential, or missing fact must come from the user.",
 				"Make one clear recommendation. Explain the decisive reasons, rejected alternatives, uncertainty, and calibrated confidence.",
 				"Return only JSON: {\"completedCriteria\":[0],\"needsUser\":false,\"nextAction\":\"...\",\"recommendation\":\"...\",\"reasons\":[\"...\"],\"rejectedAlternatives\":[{\"option\":\"...\",\"reason\":\"...\"}],\"uncertainty\":\"...\",\"confidence\":0.8}.",
@@ -120,6 +138,10 @@ var TaskJudgeService = class {
 			const modelError = getAssistantMessageErrorReason(response);
 			if (modelError) throw new Error(modelError);
 			const decision = parseTaskJudgeDecision(extractAssistantText(response.content), criteria.length);
+			if (!proof.allowed || proof.workspace && (!proof.revision || await readWorkspaceRevision(proof.workspace) !== proof.revision)) {
+				decision.completedCriteria = [];
+				decision.judgment.reasons.unshift("Current workspace verification is missing, failed or stale.");
+			}
 			const now = Date.now();
 			const evidence = decision.completedCriteria.map((index) => ({
 				kind: "state",
@@ -181,4 +203,4 @@ var TaskJudgeService = class {
 	}
 };
 //#endregion
-export { TaskJudgeService, parseTaskJudgeDecision };
+export { TaskJudgeService, codingCompletionEvidence, parseTaskJudgeDecision };

package/dist/src/agent/tools/delegate-tool.d.ts

@@ -15,7 +15,6 @@ export interface DelegateToolDeps {
         to?: string;
         threadId?: string | number;
     } | null;
-    hookRunner?: import('../../extensions/index.js').ExtensionHookRunner;
     toolExecutorConfig?: Partial<import('./executor.js').ToolExecutorConfig>;
     /**
      * Construct the child agent's tool set. Injected by `AgentToolsFactory` so
@@ -24,4 +23,6 @@ export interface DelegateToolDeps {
      */
     buildChildTools: (opts: BuildChildToolsOptions) => AgentTool<any, any>[];
 }
+export declare const INSPECTION_TOOLS: Set<string>;
+export declare function delegateToolNames(mode: 'inspect' | 'review' | 'implement', requested?: string[]): string[];
 export declare function createDelegateTool(deps: DelegateToolDeps): AgentTool;